Text-Enriched Hierarchical Graph Anomaly Detection with Uncertainty Detection

Author(s):

Xiayan Ji, John Richardson, Adrienne Raglin, Insup Lee, and Oleg Sokolsky

Abstract:

Detecting anomalous nodes in hierarchical social networks is crucial for preventing corporate fraud and conducting forensic analysis of criminal organizations. Traditional approaches rely on network analysis but overlook rich textual interactions from emails, web content, and social media. While graph neural networks show promise, they often fail to capture hierarchical structures crucial for anomaly detection. Moreover, existing methods lack robust uncertainty quantification, essential for high-stakes decision-making. We propose a novel approach integrating lightweight language models to extract textual edge embeddings, transforming them into node embeddings via convolution operations. A hyperbolic graph convolutional network models the latent hierarchy of social networks, leveraging hyperbolic space for improved hierarchical representation. Additionally, we quantify uncertainty by calibrating thresholds on node anomaly scores to ensure reliable detection. Evaluated on the real-world Enron fraud dataset and an in-house synthetic criminal network dataset, our method achieves performance comparable to large language models while significantly reducing computational overhead.